Overview
OpenAI Decisions API is an endpoint that only answers bounded questions (POST /v1/decisions), generally available since 22 October 2026. Instead of writing prose it returns one of three typed results: a predicate gives a probability from 0 to 1 that a condition holds, a choice picks one option from a set the developer defines and returns per-option probabilities with a confidence value, and a score rates an input against an ordered rubric of 2 to 10 levels as a probability-weighted average.
GPT-6 Luna is the only supported model, and OpenAI says the same question through this endpoint runs up to 10 times faster than through the Responses API. General availability raised the per-request limit from 200 to 400 questions and added prompt caching, inference residency in the US and EU, native JSON inputs, images inside question instructions, optional zero data retention and native GPT-Live integration so an app can decide while a transcript is still arriving.
Billing counts input only: $0.10 per million input tokens and $0.01 per million cached input tokens, with no charge for output tokens.
Key Features
- Three typed outputs: A predicate returns a 0 to 1 probability, a choice returns an option with per-option probabilities and a confidence value, and a score returns a weighted average across 2 to 10 ordered levels
- Up to 400 questions per request: The limit doubled from 200 when the API went generally available on 22 October 2026, and independent questions can share the same evidence
- Input-only pricing: $0.10 per million input tokens and $0.01 per million cached input tokens, with no output-token charge
- Text, image and JSON input: Accepts native JSON, conversation history and function-call records, with images placed inside question instructions
- GPT-Live voice integration: Decisions can run while a live transcript is still arriving, which suits voice apps choosing a next action or a route
- Regions and compliance: Inference residency in the US and Europe, with zero data retention available to eligible customers
Use Cases
- Content moderation and image policy checks
- Routing support tickets and scoring urgency
- Picking the next action for an agent
- Gating and grading risky instructions
- Verifying fields in forms and documents
- Real-time judgement inside voice apps
Pros
- Typed output means the app can act on the answer without parsing prose
- Input-only billing with free output keeps high-volume short calls predictable
- Up to 400 questions per request lets one piece of evidence answer several things
- Image and native JSON support widen the kinds of input it can judge
- US and EU inference residency gives a clear compliance path
- Native GPT-Live integration enables real-time decisions in voice flows
Pricing
Pricing is $0.10 per million input tokens and $0.01 per million cached input tokens, with no output-token charge and cache writes billed separately. Regional processing and long-context multipliers may apply, so check the live rate card before committing.
Summary
The Decisions API turns classification into a product of its own: the developer states the question and the answers it may return, and the endpoint hands back a probability, an option or a score. For apps that make several judgements per interaction, that beats asking a chat model to write a paragraph and then parsing it, and the boundary is just as clear, since it only answers questions that have already been framed.
Version History
- Decisions API goes generally available with typed answers (2026-10-22): OpenAI moved the Decisions API out of public beta on 22 October 2026. The POST /v1/decisions endpoint supports only GPT-6 Luna and returns three typed results: predicate, choice and score. The per-request question limit rose from 200 to 400, and general availability added prompt caching, inference residency in the US and EU, native JSON input, images inside question instructions, GPT-Live integration and zero data retention. Billing counts input at $0.10 per million tokens and $0.01 for cached input, with output free.